Train door system anomaly identification method based on deviation game and dynamic comparison

Abstract

The present application relates to the technical field of urban rail transit, in particular to a train door system anomaly identification method based on deviation game and dynamic comparison, which can learn the reasonable fluctuation range of normal working condition signals, effectively distinguish between normal signals with noise and fault signals, and alleviate the occupation of storage space and computing resources by value feature data; the present application can learn the reasonable fluctuation range of normal working condition signals, effectively distinguish between normal signals with noise and fault signals, and improve the anomaly state identification precision, through continuous compression of the occupation of storage space and computing resources by low-value feature data, the model capacity and detection performance are optimized cooperatively.

Description

Technical Field

[0001] This invention relates to the field of urban rail transit technology, specifically to an anomaly identification method for train door systems based on offset game theory and dynamic comparison. Background Technology

[0002] As a critical component of railway trains, the train door system is the most frequently operated, heaviest-loaded, and most directly interacting part of the train with passengers. Throughout its service life, it requires repeated opening and closing. Coupled with external factors such as environmental temperature changes, vibration, and shock, the system's performance gradually declines, eventually leading to malfunctions. The stability of its operational status has a significant impact on normal train operation and passenger safety. Its failures and malfunctions can easily cause service interruptions, operational delays, and other consequences, leading to a significant increase in line operation and maintenance costs. Therefore, conducting anomaly detection of the train door system and performing timely fault identification and preventative maintenance are core technical measures for strengthening the safety of train operations.

[0003] Deep learning models are widely used in the field of anomaly detection in train door systems due to their high accuracy and advantages in overcoming the limitations of traditional neural networks, such as long training time and slow convergence speed. Currently, deep learning-based anomaly detection methods are mainly divided into three categories: supervised learning, semi-supervised learning, and unsupervised learning. In actual rail transit operation scenarios, trains operate under normal conditions most of the time, resulting in a severe shortage of abnormal samples, making it impossible to cover all possible fault modes. Unsupervised learning methods only require normal operating condition data during training to identify abnormal states deviating from the normal pattern, overcoming the difficulty of acquiring abnormal data. Therefore, they have stronger engineering applicability in anomaly detection for train door systems. However, due to the influence of external noise such as environmental vibration, the signals collected by multi-source sensors have irregular and slight deviations from the signal trajectories under normal operating conditions. This causes traditional unsupervised learning methods to easily confuse noisy normal signals with early fault signals, misjudging noisy normal signals as abnormal signals, thus generating a large number of invalid warnings. Furthermore, the increase in the number and dimensions of multi-source monitoring signals acquired by different types of sensors leads to a linear increase in the storage overhead and computational complexity of traditional unsupervised learning methods. This limits the engineering deployment and long-term stable operation of traditional unsupervised learning methods in edge or online monitoring scenarios.

[0004] Therefore, how to effectively distinguish between noisy normal signals and fault signals using unsupervised learning methods, reduce the storage redundancy and computational overhead caused by multi-source high-frequency monitoring data, and improve the anomaly identification performance of train door systems are technical problems that urgently need to be solved in this field. Summary of the Invention

[0005] In view of the above-mentioned issues, this invention provides an anomaly identification method for train door systems based on offset game theory and dynamic comparison. Noise is progressively added to the original normal operating condition gate system signal samples through a Gaussian diffusion process to generate noisy comparison samples. An offset game network and a Gaussian diffusion-constrained loss function are used to constrain the feature distance between the original normal samples and the comparison samples, enabling the model to learn the reasonable fluctuation range of the signal under normal operating conditions and effectively distinguish between noisy normal signals and fault signals. By constructing a competitive memory unit of fixed capacity, low-weight redundant samples are dynamically forgotten and eliminated according to the importance weight of the input samples, continuously compressing the storage and computing resource consumption of low-value feature data, thus achieving synergistic optimization of model capacity and detection performance.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: An anomaly detection method for train door systems based on offset game theory and dynamic comparison includes the following steps: S1. Obtain monitoring data of the train door system; S2. Extract the statistical features of the train door system signals collected under normal operating conditions, and combine the statistical features with the number of iteration steps of the Gaussian diffusion process to obtain the iterative statistical features; S3. Use a bi-branch nonlinear mapping to fit the mean and standard deviation constraint variables of the Gaussian diffusion process; S4. Under the constraints of the mean and standard deviation of the Gaussian diffusion process, noise is gradually injected into the original signal through the Gaussian diffusion process to generate a noisy comparison sample. S5. Construct an offset game network to extract feature vectors from normal samples and noisy contrast samples, respectively. S6. Train the offset game network using the joint loss function; S7. Construct a competitive memory unit with a fixed capacity to store noisy comparison samples. During the training process, calculate the importance weight of each sample in the memory unit, eliminate low-weight redundant samples and replace them with high-weight samples. S8. Input the real-time acquired gate system signals into the trained model to obtain scores, using 3 The criteria determine whether the score exceeds the abnormal threshold, thereby outputting the health status of the gate system and achieving early warning.

[0007] Preferably, in S1, the monitoring data of the train door system is defined as follows: ,in N Number of sensor types L To monitor the length of the time window, , For the first i Time-series data monitored by each sensor.

[0008] Preferably, in S2, the statistical characteristics of the train door system signals acquired under normal operating conditions include the average value, standard deviation, maximum value, and minimum value, which are represented in vector form as follows: In the formula, and Signals The mean and standard deviation; and They represent signals respectively The maximum and minimum values; Indicates signal The statistical characteristics are then used. The formula for calculating the iterative statistical characteristics, combining the statistical characteristics with the number of iteration steps, is as follows: In the formula, T This represents the total number of steps in the Gaussian diffusion. Indicates signal The t The statistical feature vector of step iteration, .

[0009] Preferably, in S3, the formula for calculating the mean and standard deviation constraint variables of the Gaussian diffusion process using a bi-branch nonlinear mapping is as follows: In the formula, ReLU tanh and Softplus For activation functions; , and These are the statistical characteristics of the transitional state. Mean constraint variables and standard deviation constraint variables The corresponding linear mapping weight matrix; , and These are the statistical characteristics of the transitional state. Mean constraint variables and standard deviation constraint variables The corresponding linear mapping bias term; and These represent the sensors. i Maximum baseline offset and maximum inherent fluctuation amplitude under normal operating conditions; Indicates signal No. t Iteratively analyze the transition states of the statistical feature vectors; Indicates signal No.t Mean constraint variables for each iteration; Indicates signal No. t The standard deviation constraint variable for each iteration.

[0010] Preferably, in S4, under the constraints of the iterative mean and standard deviation of the Gaussian diffusion process, the formula for injecting noise into the original signal through a progressive Gaussian diffusion process to generate a noisy comparison sample is as follows: in, For the first t Noise scheduling rate of Gaussian spread step; Indicates a standard Gaussian distribution; Indicates signal No. t Mean constraint variables for each iteration; Indicates signal No. t Standard deviation constraint variable for each iteration; This represents a random transformation matrix sampled from a Gaussian distribution; This represents a random vector drawn from a Gaussian distribution; Indicates the process t Signal obtained by Gaussian diffusion The corresponding noisy comparison samples, .

[0011] Preferably, in S5, the offset game network includes: S51, Signals for train door system under normal operating conditions And a noisy comparison sample generated through a Gaussian diffusion process. After normalization and feature mapping, feature vectors of uniform dimension are obtained: In the formula, and They represent and The feature vector with uniform dimensions obtained after preprocessing; and They represent and The mean; and They represent and Standard deviation; and They represent respectively to The weight matrix and bias terms of the feature mapping operation performed; and They represent respectively to The weights and biases of the feature mapping operations performed; S52, Feature Vector and Simultaneously passing through two convolutional layers, the output is an encoded feature map: In the formula, Represents the ReLU activation function; and These represent the first and second convolutional operations on the encoding side of the offset game network, respectively. and This indicates the kernel size of the corresponding convolutional layer. and This indicates the number of convolutional kernels in the corresponding convolutional layer; and express and Encoded feature map output from the encoding side of the offset game network; S53, will and Synchronous input gated recurrent unit, outputting feature maps with strong representational capabilities: In the formula, Indicates a gated loop unit. cell For the hidden layer dimension; and They represent and A feature map with strong representational capabilities output by the gated loop unit; S54, and Simultaneously, the signal passes through two decoding convolutional layers. The output of the second convolutional layer on the encoding side is added to the input of the first convolutional layer on the decoding side, and then fed into the first convolutional layer on the decoding side. Similarly, the output of the first convolutional layer on the encoding side is added to the input of the second convolutional layer on the decoding side, and then fed into the second convolutional layer on the decoding side. In the formula, This indicates an addition operation; and They represent and The feature vector with uniform dimensions obtained after preprocessing; This represents the feature vector obtained by adding the output of the first convolutional layer on the encoding side to the input of the second convolutional layer on the decoding side. The feature vector representing the sum of the output of the second convolutional layer on the encoding side and the input of the first convolutional layer on the decoding side is calculated using the following formula: In the formula, This indicates an addition operation; and express and Encoded feature map output from the encoding side of the offset game network; This represents the first convolutional operation on the decoding side of the offset game network; S55, Pass the last convolutional layer on the decoding side. The features are mapped to output a fixed-dimensional representation vector for normal samples and a representation vector for noisy contrast samples: In the formula, This represents the second convolutional operation on the decoding side of the offset game network; and They represent and The decoded feature map output by the decoder side of the offset game network.

[0012] Preferably, in S6, a joint loss function is designed to reduce the model's sensitivity to normal operating noise such as environmental vibration and to improve the model's accuracy in identifying abnormalities in the train door system.

[0013] Preferably, S6 includes: S61, constructing a biased game loss function. : In the formula, N Number of sensor types For decoding feature map dimensions; For temperature parameters; and They represent and The l One feature point; Represents the cosine similarity function; S62. Constructing the Gaussian diffusion-constrained loss function : In the formula, Indicates signal No. t Mean constraint variables for each iteration; Indicates signal No. t Standard deviation constraint variable for each iteration; N Number of sensor types T This represents the total number of steps in the Gaussian diffusion. S63. Constructing a binary classification cross-entropy loss function : In the formula, B This refers to the training batch size; For the first in a batch b The true label of each sample; For the first in a batch b The predicted probability of anomalies for each sample; S64. The joint loss function is obtained by weighting the offset game loss function, the Gaussian diffusion constraint loss function, and the binary classification cross-entropy loss function: In the formula, , and These are hyperparameters used to adjust the weights of the corresponding losses.

[0014] Preferably, in S7, during the training process, under the statistical characteristic constraints of the door system signals under normal operating conditions, noisy contrast samples generated through a progressive Gaussian diffusion process are input into a system with a fixed capacity. C The importance weight is calculated in the competitive memory units using the following formula: In the formula, N This represents the number of sensor types, i.e., the number of multi-source feature channels in the sample. For decoding feature dimensions; and They represent and The l One feature point; Represents the cosine similarity function; express for Importance weights.

[0015] Preferably, in S8, 3 The criteria determine whether the score of the input sample exceeds the abnormal threshold.

[0016] Preferably, the design 3 The calculation process for the criteria includes: S81. Collect and score the following door system signal samples under normal operating conditions: In the formula, F This represents the set of scores for samples under normal operating conditions. Indicates the first m Score of a normal operating condition sample ; S82, Calculation F Mean of scores of all samples and standard deviation : S83, Use Establish abnormal thresholds for the train door system: In the formula, This represents the upper boundary of the abnormal threshold. The lower boundary of the anomaly threshold is where the input sample score is within. If the sample is internal, it is considered a normal sample; otherwise, it is considered an abnormal sample.

[0017] As can be seen from the above technical solution, compared with the prior art, the present invention has the following beneficial effects: This invention addresses the anomaly detection needs of train door system operation data. It can learn the reasonable fluctuation range of signals under normal operating conditions, effectively distinguish between noisy normal signals and fault signals, and improve the accuracy of anomaly identification. By continuously compressing the storage space and computing resources occupied by low-value feature data, it achieves synergistic optimization of model capacity and detection performance, and has the following technical features: 1) Noise is gradually added to the original normal operating condition gate system signal samples through a Gaussian diffusion process to generate noisy comparison samples. The feature distance between the original normal samples and the comparison samples is constrained by an offset game network and a Gaussian diffusion constraint loss function, so that the model can learn the reasonable fluctuation range of the signal under normal operating conditions and effectively distinguish between noisy normal signals and fault signals; 2) Construct competitive memory units of fixed size, dynamically forget and eliminate low-weight redundant samples according to the importance weight of input samples, continuously compress the storage and computing resources occupied by low-value feature data, and realize the synergistic optimization of model capacity and detection performance. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0019] Figure 1A flowchart of an anomaly identification method for a train door system based on offset game theory and dynamic comparison provided by the present invention; Figure 2 The basic structure diagram of the gate system anomaly recognition network provided by the present invention is shown. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] See Figures 1 to 2 This invention discloses an anomaly identification method for train door systems based on offset game theory and dynamic comparison, comprising the following steps: S1. Obtain monitoring data of the train door system; S2. Extract the statistical features of the train door system signals collected under normal operating conditions, and combine the statistical features with the number of iteration steps of the Gaussian diffusion process to obtain the iterative statistical features; S3. Use a bi-branch nonlinear mapping to fit the mean and standard deviation constraint variables of the Gaussian diffusion process; S4. Under the constraints of the mean and standard deviation of the Gaussian diffusion process, noise is gradually injected into the original signal through the Gaussian diffusion process to generate a noisy comparison sample. S5. Construct an offset game network to extract feature vectors from normal samples and noisy contrast samples, respectively. S6. Train the offset game network using the joint loss function; S7. Construct a competitive memory unit with a fixed capacity to store noisy comparison samples. During the training process, calculate the importance weight of each sample in the memory unit, eliminate low-weight redundant samples and replace them with high-weight samples. S8. Input the real-time acquired gate system signals into the trained model to obtain scores, using 3 The criteria determine whether the score exceeds the abnormal threshold, thereby outputting the health status of the gate system and achieving early warning.

[0022] It should be noted that the monitoring data mentioned above refers to time series data such as current, displacement, and speed of the door system drive motor; the statistical characteristics mentioned refer to the state quantities and statistics collected during the operation of the door system, such as door opening and closing time and maximum current; the raw signals mentioned refer to the unprocessed monitoring data in time series form; and the normal samples mentioned refer to data samples collected when the door system is operating under normal conditions.

[0023] Preferably, in S1, the monitoring data of the train door system is defined as follows: ,in N Number of sensor types L To monitor the length of the time window, , For the first i Time-series data monitored by each sensor.

[0024] Preferably, in S2, the statistical characteristics of the train door system signals acquired under normal operating conditions include the average value, standard deviation, maximum value, and minimum value, which are represented in vector form as follows: In the formula, and Signals The mean and standard deviation; and They represent signals respectively The maximum and minimum values; Indicates signal The statistical characteristics are then used. The formula for calculating the iterative statistical characteristics, combining the statistical characteristics with the number of iteration steps, is as follows: In the formula, T This represents the total number of steps in the Gaussian diffusion. Indicates signal The t The statistical feature vector of step iteration, .

[0025] Preferably, in S3, the formula for calculating the mean and standard deviation constraint variables of the Gaussian diffusion process using a bi-branch nonlinear mapping is as follows: In the formula, ReLU tanh and Softplus For activation functions; , and These are the statistical characteristics of the transitional state. Mean constraint variables and standard deviation constraint variables The corresponding linear mapping weight matrix; , and These are the statistical characteristics of the transitional state. Mean constraint variables and standard deviation constraint variables The corresponding linear mapping bias term; and These represent the sensors. i Maximum baseline offset and maximum inherent fluctuation amplitude under normal operating conditions; Indicates signal No. t Iteratively analyze the transition states of the statistical feature vectors; Indicates signal No. t Mean constraint variables for each iteration; Indicates signal No. t The standard deviation constraint variable for each iteration.

[0026] Preferably, in S4, under the constraints of the iterative mean and standard deviation of the Gaussian diffusion process, the formula for injecting noise into the original signal through a progressive Gaussian diffusion process to generate a noisy comparison sample is as follows: in, For the first t Noise scheduling rate of Gaussian spread step; Indicates a standard Gaussian distribution; Indicates signal No. t Mean constraint variables for each iteration; Indicates signal No. t Standard deviation constraint variable for each iteration; This represents a random transformation matrix sampled from a Gaussian distribution; This represents a random vector drawn from a Gaussian distribution; Indicates the process t Signal obtained by Gaussian diffusion The corresponding noisy comparison samples, .

[0027] Preferably, in S5, the offset game network includes: S51, Signals for train door system under normal operating conditions And a noisy comparison sample generated through a Gaussian diffusion process. After normalization and feature mapping, feature vectors of uniform dimension are obtained: In the formula, and They represent and The feature vector with uniform dimensions obtained after preprocessing; and They represent and The mean; and They represent and Standard deviation; and They represent respectively to The weight matrix and bias terms of the feature mapping operation performed; and They represent respectively to The weights and biases of the feature mapping operations performed; S52, Feature Vector and Simultaneously passing through two convolutional layers, the output is an encoded feature map: In the formula, Represents the ReLU activation function; and These represent the first and second convolutional operations on the encoding side of the offset game network, respectively. and This indicates the kernel size of the corresponding convolutional layer. and This indicates the number of convolutional kernels in the corresponding convolutional layer; and express and Encoded feature map output from the encoding side of the offset game network; S53, will and Synchronous input gated recurrent unit, outputting feature maps with strong representational capabilities: In the formula, Indicates a gated loop unit. cell For the hidden layer dimension; and They represent and A feature map with strong representational capabilities output by the gated loop unit; S54, and Simultaneously, the signal passes through two decoding convolutional layers. The output of the second convolutional layer on the encoding side is added to the input of the first convolutional layer on the decoding side, and then fed into the first convolutional layer on the decoding side. Similarly, the output of the first convolutional layer on the encoding side is added to the input of the second convolutional layer on the decoding side, and then fed into the second convolutional layer on the decoding side. In the formula, This indicates an addition operation; and They represent and The feature vector with uniform dimensions obtained after preprocessing; This represents the feature vector obtained by adding the output of the first convolutional layer on the encoding side to the input of the second convolutional layer on the decoding side. The feature vector representing the sum of the output of the second convolutional layer on the encoding side and the input of the first convolutional layer on the decoding side is calculated using the following formula: In the formula, This indicates an addition operation; and express and Encoded feature map output from the encoding side of the offset game network; This represents the first convolutional operation on the decoding side of the offset game network; S55, Pass the last convolutional layer on the decoding side. The features are mapped to output a fixed-dimensional representation vector for normal samples and a representation vector for noisy contrast samples: In the formula, This represents the second convolutional operation on the decoding side of the offset game network; and They represent and The decoded feature map output by the decoder side of the offset game network.

[0028] Preferably, in S6, a joint loss function is designed to reduce the model's sensitivity to normal operating noise such as environmental vibration and to improve the model's accuracy in identifying abnormalities in the train door system.

[0029] Preferably, S6 includes: S61, constructing a biased game loss function. : In the formula, N Number of sensor types For decoding feature map dimensions; For temperature parameters; and They represent and The l One feature point; Represents the cosine similarity function; S62. Constructing the Gaussian diffusion-constrained loss function : In the formula, Indicates signal No. t Mean constraint variables for each iteration; Indicates signal No. t Standard deviation constraint variable for each iteration; N Number of sensor types T This represents the total number of steps in the Gaussian diffusion. S63. Constructing a binary classification cross-entropy loss function : In the formula, B This refers to the training batch size; For the first in a batch b The true label of each sample; For the first in a batch b The predicted probability of anomalies for each sample; S64. The joint loss function is obtained by weighting the offset game loss function, the Gaussian diffusion constraint loss function, and the binary classification cross-entropy loss function: In the formula, , and These are hyperparameters used to adjust the weights of the corresponding losses.

[0030] Preferably, in S7, during the training process, under the statistical characteristic constraints of the door system signals under normal operating conditions, noisy contrast samples generated through a progressive Gaussian diffusion process are input into a system with a fixed capacity. C The importance weight is calculated in the competitive memory units using the following formula: In the formula, N This represents the number of sensor types, i.e., the number of multi-source feature channels in the sample. For decoding feature dimensions; and They represent and The l One feature point; Represents the cosine similarity function; express for Importance weights.

[0031] Preferably, in S8, 3 The criteria determine whether the score of the input sample exceeds the abnormal threshold.

[0032] Preferably, the design 3 The calculation process for the criteria includes: S81. Collect and score the following door system signal samples under normal operating conditions: In the formula, F This represents the set of scores for samples under normal operating conditions. Indicates the first m Score of a normal operating condition sample ; S82, Calculation F Mean of scores of all samples and standard deviation : S83, Use Establish abnormal thresholds for the train door system: In the formula, This represents the upper boundary of the abnormal threshold. The lower boundary of the anomaly threshold is where the input sample score is within. If the sample is internal, it is considered a normal sample; otherwise, it is considered an abnormal sample.

[0033] Next, taking the rail vehicle door system as an example, the effectiveness of the method of the present invention was verified by collecting multi-source sensor data from the door systems of in-service railway vehicles on urban rail transit lines in China and simulating abnormal signals from actual vehicles. The experiment selected data from six door systems as the training set, generating noisy comparison samples. Data from another two door systems and eight sets of simulated abnormal signals were selected as the test set. The specific data design is shown in Table 1. Table 1. Division of Experimental Data

[0034] During data acquisition, multi-source sensors were used to collect three types of signals: motor displacement, motor speed, and motor current, during the complete opening and closing process of the door system. The sampling frequency was 100Hz. Four fault states were simulated by locally attaching tape to the sliding rail of the door system: abnormal sealing size, abnormal door opening and closing time, abnormal local resistance, and abnormal global resistance. The time taken for one door opening and closing was 3.5s, and the signal data length was 350. To verify the robustness and superiority of the proposed door system anomaly identification method, this invention was compared with three more advanced door system temporal anomaly identification methods, including: anomaly detection model based on Minimal Memory Enhancement Reconstruction (DiMER), anomaly identification method based on Importance Generative Contrastive Learning (IGCL), and anomaly identification method based on Contrastive Representation Learning (CARL), corresponding to methods 1 to 4 respectively. All comparison methods were trained using the same hyperparameters and network structure. Accuracy, recall, and macro-average F1 score were selected as model performance evaluation indicators. The comparative experimental results are shown in Table 2. Table 2. Performance Comparison of Four Methods

[0035] The proposed method achieves the best performance in all three core evaluation metrics, demonstrating the best anomaly identification performance. Specifically, the accuracy, recall, and macro F1 score are improved by an average of approximately 8%, 11%, and 7% respectively compared to the best comparison method, which fully verifies the advantages of the proposed method in anomaly identification of gate systems.

[0036] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0037] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for anomaly identification in a train door system using offset game theory and dynamic comparison, characterized in that, Includes the following steps: S1. Obtain monitoring data of the train door system; S2. Extract the statistical features of the train door system signals collected under normal operating conditions, and combine the statistical features with the number of iterations set for the Gaussian diffusion process to obtain the iterated statistical features. S3. Use a bi-branch nonlinear mapping to fit the mean and standard deviation constraint variables of the Gaussian diffusion process; S4. Under the constraints of the mean and standard deviation of the Gaussian diffusion process, noise is gradually injected into the original signal through the Gaussian diffusion process to generate a noisy comparison sample. S5. Construct an offset game network to extract feature vectors from normal samples and noisy contrast samples, respectively. S6. Train the offset game network using the joint loss function; S7. Construct a fixed-capacity competitive memory unit to store noisy comparison samples. During training, calculate the importance weight of each sample in the memory unit, eliminate low-weight redundant samples and retain high-weight samples. S8. Input the real-time acquired gate system signals into the trained model to obtain scores, using 3 The criteria determine whether the score exceeds the abnormal threshold, thereby outputting the health status of the gate system and achieving early warning.

2. The train door system anomaly identification method based on offset game theory and dynamic comparison according to claim 1, characterized in that, In S1, the monitoring data of the train door system is defined as follows: ,in N Number of sensor types L To monitor the length of the time window, , For the first i Time-series data monitored by each sensor.

3. The train door system anomaly identification method based on offset game theory and dynamic comparison according to claim 2, characterized in that, In S2, the statistical characteristics of the train door system signals collected under normal operating conditions include the average value, standard deviation, maximum value, and minimum value, which are represented in vector form as follows: In the formula, and Signals The mean and standard deviation; and They represent signals respectively The maximum and minimum values; Indicates signal The statistical characteristics, combined with the number of iteration steps, yield the following formula for calculating the iterative statistical characteristics: In the formula, T This represents the total number of steps in the Gaussian diffusion. Indicates signal The t Statistical feature vectors obtained through step-by-step iteration. .

4. The train door system anomaly identification method based on offset game theory and dynamic comparison according to claim 3, characterized in that, In S3, the formula for calculating the mean and standard deviation constraint variables of the Gaussian diffusion process using a bi-branch nonlinear mapping is as follows: In the formula, ReLU tanh and Softplus For activation functions; , and These are the statistical characteristics of the transitional state. Mean constraint variables and standard deviation constraint variables The corresponding linear mapping weight matrix; , and These are the statistical characteristics of the transitional state. Mean constraint variables and standard deviation constraint variables The corresponding linear mapping bias term; and These represent the sensors. i Maximum baseline offset and maximum inherent fluctuation amplitude under normal operating conditions; Indicates signal No. t Iteratively analyze the transition states of the statistical feature vectors; Indicates signal No. t Mean constraint variables for each iteration; Indicates signal No. t The standard deviation constraint variable for each iteration.

5. The train door system anomaly identification method based on offset game theory and dynamic comparison according to claim 4, characterized in that, In step S4, under the constraints of the iterative mean and standard deviation of the Gaussian diffusion process, noise is injected into the original signal through a progressive Gaussian diffusion process to generate a noisy comparison sample, as follows: in, For the first t Noise scheduling rate of Gaussian spread step; Indicates a standard Gaussian distribution; Indicates signal No. t Mean constraint variables for each iteration; Indicates signal No. t Standard deviation constraint variable for each iteration; This represents a random transformation matrix sampled from a Gaussian distribution; This represents a random vector drawn from a Gaussian distribution; Indicates the process t Signal obtained by Gaussian diffusion The corresponding noisy comparison samples, .

6. The train door system anomaly identification method based on offset game theory and dynamic comparison according to claim 5, characterized in that, S5 includes: S51, Signals for train door system under normal operating conditions And a noisy comparison sample generated through a Gaussian diffusion process. After normalization and feature mapping, feature vectors of uniform dimension are obtained: In the formula, and They represent and The feature vector with uniform dimensions obtained after preprocessing; and They represent and The mean; and They represent and Standard deviation; and They represent respectively to The weight matrix and bias terms of the feature mapping operation performed; and They represent respectively to The weights and biases of the feature mapping operations performed; S52, Feature Vector and Simultaneously passing through two convolutional layers, the output is an encoded feature map: In the formula, Represents the ReLU activation function; and These represent the first and second convolutional operations on the encoding side of the offset game network, respectively. and This indicates the kernel size of the corresponding convolutional layer. and This indicates the number of convolutional kernels in the corresponding convolutional layer; and express and Encoded feature map output from the encoding side of the offset game network; S53, will and Synchronous input gated recurrent unit, outputting feature maps with strong representational capabilities: In the formula, Indicates a gated loop unit. cell For the hidden layer dimension; and They represent and A feature map with strong representational capabilities output by the gated loop unit; S54, and Simultaneously, the signal passes through two decoding convolutional layers. The output of the second convolutional layer on the encoding side is added to the input of the first convolutional layer on the decoding side, and then fed into the first convolutional layer on the decoding side. Similarly, the output of the first convolutional layer on the encoding side is added to the input of the second convolutional layer on the decoding side, and then fed into the second convolutional layer on the decoding side. In the formula, This indicates an addition operation; and They represent and The feature vector with uniform dimensions obtained after preprocessing; This represents the feature vector obtained by adding the output of the first convolutional layer on the encoding side to the input of the second convolutional layer on the decoding side. The feature vector representing the sum of the output of the second convolutional layer on the encoding side and the input of the first convolutional layer on the decoding side is calculated using the following formula: In the formula, This indicates an addition operation; and express and Encoded feature map output from the encoding side of the offset game network; This represents the first convolutional operation on the decoding side of the offset game network; S55, Pass the last convolutional layer on the decoding side. The features are mapped to output a fixed-dimensional representation vector for normal samples and a representation vector for noisy contrast samples: In the formula, This represents the second convolutional operation on the decoding side of the offset game network; and They represent and The decoded feature map output by the decoder side of the offset game network.

7. The train door system anomaly identification method based on offset game theory and dynamic comparison according to claim 6, characterized in that, S6 includes: S61. Constructing the offset game loss function : In the formula, N Number of sensor types For decoding feature map dimensions; For temperature parameters; and They represent and The l One feature point; Represents the cosine similarity function; S62. Constructing the Gaussian diffusion-constrained loss function : In the formula, Indicates signal No. t Mean constraint variables for each iteration; Indicates signal No. t Standard deviation constraint variable for each iteration; N Number of sensor types T This represents the total number of steps in the Gaussian diffusion. S63. Constructing a binary classification cross-entropy loss function : In the formula, B This refers to the training batch size; For the first in a batch b The true label of each sample; For the first in a batch b The predicted probability of anomalies for each sample; S64. The joint loss function is obtained by weighting the offset game loss function, the Gaussian diffusion constraint loss function, and the binary classification cross-entropy loss function: In the formula, , and These are hyperparameters used to adjust the weights of the corresponding losses.

8. The train door system anomaly identification method based on offset game theory and dynamic comparison according to claim 7, characterized in that, In step S7, during the training process, under the statistical characteristic constraints of the door system signals under normal operating conditions, noisy contrast samples generated through a progressive Gaussian diffusion process are input into a database with a fixed capacity. C The importance weight is calculated in the competitive memory units using the following formula: In the formula, N This represents the number of sensor types, i.e., the number of multi-source feature channels in the sample. For decoding feature dimensions; and They represent and The l One feature point; Represents the cosine similarity function; express for Importance weights.

9. The train door system anomaly identification method based on offset game theory and dynamic comparison according to claim 8, characterized in that, S8 includes: S81. Collect and score the following door system signal samples under normal operating conditions: In the formula, F This represents the set of scores for samples under normal operating conditions. Indicates the first m Score of a normal operating condition sample ; S82, Calculation F Mean of scores of all samples and standard deviation : S83, Use Establish abnormal thresholds for the train door system: In the formula, This represents the upper boundary of the abnormal threshold. The lower boundary of the anomaly threshold is where the input sample score is within. If the sample is internal, it is considered a normal sample; otherwise, it is considered an abnormal sample.

Images